Message Passing Neural Networks for Partial Charge Assignment in Metal-Organic Frameworks

Virtual screenings of thousands of metal-organic frameworks (MOFs) can accelerate and reduce the cost of discovering MOFs for their applications in gas storage, separation/purification, and sensing. In molecular mechanics simulations of gas adsorption/diffusion in MOFs, the adsorbate-MOF electrostatic interaction is typically modeled by placing partial charges on the atoms of the MOF. It is critical to develop computationally inexpensive methods to assign atomic partial charges to a MOF that accurately reproduce the electrostatic potential in its pores. Herein, we design and train a message passing neural network to predict the atomic partial charges on MOFs under a charge neutral constraint. A set of ca. 2266 MOFs labeled with high-fidelity partial charges, derived from electronic structure calculations, serve as training examples. In an end-to-end manner, from the crystal graph representing the MOF, our message passing neural network machine-learns features of the local bonding environments of each atom and learns to predict partial charges from these features. Our trained message passing neural network can assign high-fidelity partial point charges to MOFs with orders of magnitude lower computational cost, thereby enhancing the accuracy of virtual screenings of MOFs for their adsorption-based applications.